Classification of Apple Growing Region by Vis-Near Infrared Spectroscopy Coupled with Machine Learning and Deep Learning Methods

This study explores how effective VIS-NIR spectroscopy, combined with machine learning and deep learning techniques, is in classifying various apple varieties gathered from two distinct locations. The study analyzed the diffuse reflectance spectra of three types of Chinese apples: the Fuji apple, the Red Star apple, and the Gala apple. Each apple variety was collected in two distinct collection locations: Shandong and Shaanxi. To achieve this, a comparison between machine learning methods (partial least squares discriminant analysis (PLS-DA) and deep learning methods (artificial neural networks discriminant analysis (ANN-DA)) has been performed. The results show that the ANN-DA models outperform the PLS-DA and SVM models. For the entire sample set, the fact of having three different apple varieties makes all three algorithms the task of classifying correctly more difficult. When the two locations are classified using samples from the same variety, ANN has an accuracy in the calibration and prediction sets of 100%, except for the fuji variety, which reaches 94%. These results show that NIR spectroscopy coupled with Machine Learning or Deep Learning models is a capable method for identifying the apple growing region even for industrial-scale adoption.